Page 192 -

180 5 Neural Networks

The training set, validation set and test set errors for this experiment with the
MLP2:2:3 are 8.2%, 7.7% and 10.5%, respectively. When conducting an
experiment with small datasets andlor low dimensionality ratios, it is advisable not
to trust just one run. Instead, several runs should be performed using randomised
sets or the partition method explained in section 4.5. When performing ten runs
with the cork stoppers data, by randomly shuffling the cases, we obtained different
solutions regarding the relative value of the errors, all with small deviations from
the previously mentioned results. This is, of course, a good indication that the
neural net is not over-fitting the data. Table 5.6 summarizes the results of these ten
runs.

Table 5.6. Statistical results of ten randomised runs with the MLP2:2:3.
Training Validation Test
Average error 10.5 9.2 10.3

Standard deviation 2.8 3.4 3.9

From these results we can obtain a better idea of the attainable network
performance, and conclude that the overall error should be near 10% with
approximately 3% standard deviation.

Case 1 Case 101 Case 201 Case 301 Case 401
Case 51 Case 151 Case 251 Case 351
Figure 5.27. Predicted foetal weight (PR-FW) using an MLP3:6: 1 trained with the
back-propagation algorithm. The FW curve represents the true foetal weight
values.